Analysis Of Complex-valued FMRI Data Based On Phase Information Of ICA Components

Functional magnetic resonance imaging (fMRI) is a powerful and noninvasive tool utilized in both research and clinical arenas since the early1990s. As a type of blind source separation (BSS) that requires little prior information about the brain, ICA has become an attractive approach for finding the underlying time courses and spatial maps in the brain from the fMRI data.FMRI data are initially acquired as complex-valued image pairs including both magnitude and phase. Due to its unknown and noisy nature, the phase data are completely discarded in the vast majority of fMRI analyses. However, some previous studies show that the phase data contain useful and unique information for better understanding brain function. The inclusion of the phase provides the possibility of extracting more complete information from fMRI data. As such, this thesis proposes a novel framework for utilizing the phase information recovered by complex-valued ICA, in which the motor tapping complex-valued fMRI data are used. The main work is as follows:(1) The thesis proposes a new phase ambiguity correction scheme that solves the inherent phase scaling ambiguity found in complex-valued ICA. It is based on the time course estimates from ICA as they have much higher degrees of non-circularity than the spatial map estimates. In addition, the above-mentioned phase de-ambiguity suffers from sign ambiguity. This problem can be reliably solved when some prior information about the time course or spatial map is available. Experimental results show that this method provides better sensitivity and accuracy.(2) The thesis then proposes the concept of phase positioning by defining criteria to segment voxels in the spatial map estimates from ICA into the target and interference voxels. And presents a novel visualization technique, based on the combination of phase range and magnitude strength. Along this line, the single-subject and group phase masks are constructed and the phase masking algorithms are presented to remove the interference voxels from the individual and group spatial map estimates. In this case, the phase can be efficiently used to distinguish the target voxels from the interference voxels. In addition, the interference voxels can be removed after doing ICA without need of removing the noisy voxels prior to ICA, which can guarantee the integrity of the fMRI data. The experimental results show its high accuracy. (3) To further verify the efficiency of the proposed framework, the thesis extracts the main brain networks from the motor-tapping fMRI data. The results show the potential capability of the proposed framework in enhancing the performance of clinical studies when using complex-valued fMRI data.